Tabular foundation models are the AI category nobody built a consumer app for, and SAP has just put more than €1 billion behind them.

By Blumefield | 21 July 2026

The Deal That Closed Without a Product Launch

On 17 July, SAP confirmed it had completed the purchase of Prior Labs, the Freiburg research outfit that turned tabular foundation models from an academic curiosity into something enterprises will actually pay for. The German software group had announced the agreement in May and spent the intervening ten weeks clearing regulatory approvals. There was no keynote, no demo, no launch video. There was a two-paragraph newsbyte from Walldorf and a short post from three co-founders saying the money had landed.

The money is the story. SAP has committed more than €1 billion over four years to scale Prior Labs into what its own announcement calls a globally leading frontier AI lab. That figure puts a single European enterprise software vendor into the same conversation as the venture rounds being raised by American model labs, and it does so in a category most of the AI industry has ignored. Prior Labs was founded eighteen months ago. Terms of the acquisition itself were never disclosed, though European reporting has put the combined value comfortably north of a billion euros.

What SAP bought is not a chatbot. It is a bet that the most commercially defensible corner of artificial intelligence is the one nobody made a consumer app out of.

Why Tables Beat Chat in the Enterprise

Large language models are extraordinary at prose and mediocre at arithmetic on structured records. They have, as SAP chief technology officer Philipp Herzig put it, only a rudimentary understanding of tables, numbers and statistics. Ask a frontier LLM to predict which of forty thousand invoices will be paid late and it will produce something confident and unreliable. That gap matters because the overwhelming majority of enterprise data does not live in documents. It lives in rows and columns inside ERP systems, and SAP happens to run more of those than anyone.

Tabular foundation models attack the problem directly, and they do it with an architecture that looks nothing like a chatbot. Prior Labs' TabPFN is pretrained on synthetic data and performs in-context learning on whatever table you hand it, which means no bespoke model training per dataset, no feature engineering pipeline, no six-week data science project. The original method was published in Nature in January 2025 and showed the model beating an ensemble of tuned baselines that had been given four hours of compute, in under three seconds. TabPFN-3, released in May 2026, extends that to a million rows and returns predictions in roughly two tenths of a second.

The commercial translation is blunt. Payment delays, supplier risk, customer churn, demand forecasting and upsell scoring are all tabular prediction problems, and they are all things enterprises currently solve with expensive, brittle, one-off machine learning. Collapsing that into a single pretrained model is a genuine change in unit economics, not a marketing reframe. Hitachi Rail reports a 40 percent improvement in anomaly detection for predictive maintenance. TD's Layer 6 AI group is testing the approach for financial forecasting at scale.

SAP Already Believed This

The acquisition is not opportunistic. SAP built its own tabular model, SAP-RPT-1, before it went shopping, which is the clearest possible signal that the strategy predates the target. Herzig has been explicit that SAP concluded early on that the greatest untapped opportunity in enterprise AI was never large language models but the structured data that runs the world's businesses. Buying the leading external team in the category is what conviction looks like when you already have a house model and want to stop being second best at your own idea.

The integration path for tabular foundation models is already mapped. Prior Labs' work is intended to flow into SAP AI Core and SAP Business Data Cloud, and then into Joule, SAP's agentic layer. That last piece is where the strategic logic tightens. An agent that can act on enterprise systems is only as useful as its ability to forecast consequences, and a language model alone cannot do that on a general ledger. Pairing an agentic interface with tabular foundation models gives SAP something its rivals cannot easily assemble, because the rivals do not own the tables.

It also sits inside a wider buying spree. SAP closed its Reltio acquisition in May and its Dremio acquisition on 6 July, both aimed at unifying SAP and non-SAP data. The Prior Labs deal is the modelling layer on top of a data estate SAP has spent the year consolidating.

The Open Source Condition

The genuinely unusual term is that Prior Labs stays open. The lab keeps its brand, its leadership, its Freiburg headquarters, its Berlin and New York offices, its research agenda and its customer relationships, and it continues to publish and to release models publicly. SAP has said it is fully committed to supporting that open source strategy. TabPFN has passed four million downloads and eight thousand GitHub stars, and the model repositories remain openly distributed alongside deployments on Databricks, AWS SageMaker and Azure AI Foundry.

Corporate acquirers do not usually leave that door open when they have just paid a billion euros for tabular foundation models. The commercial instinct is to close the weights and monetise the moat, and the graveyard of acquired open source projects is well populated. SAP is calculating, plausibly, that in tabular foundation models the moat is not the weights at all. It is the data, the distribution and the enterprise trust to run predictions against a live financial system under GDPR. Give away the model, keep the deployment.

The scientific advisory board reinforces the point. Yann LeCun, Turing Award winner and executive chairman at Advanced Machine Intelligence, and Bernhard Schölkopf of the Max Planck Institute for Intelligent Systems both serve on it. Neither is the sort of name that attaches itself to a closed corporate research unit.

Europe's Cleanest Sovereignty Test

European policymakers have spent the better part of a decade trying to manufacture exactly this outcome and mostly failing. The pattern has been depressingly stable. European academic labs produce world-class research, the research spins out into a startup, the startup raises from American investors, and the talent and the intellectual property migrate west. DeepMind went to Google. The pipeline leaked at the commercial stage every time.

Here a European academic group at Freiburg published in Nature, spun out, scaled in eighteen months, and was acquired by a European incumbent that then committed a billion euros to keep it independent, in Germany, and open. As Blumefield has argued through this year's run of EU tech sovereignty announcements, the Chips Act 2.0 and the CADA package included, subsidy programmes have consistently underperformed the thing that actually works, which is a domestic buyer with a real balance sheet and a strategic reason to pay up.

The caveats are real. A billion euros over four years is roughly what a US hyperscaler spends on AI infrastructure in a fortnight. Independence inside a fifty-year-old enterprise software company is a promise that survives exactly as long as the quarterly numbers allow, and SAP's own capital allocation will be tested if enterprise AI budgets tighten. Tabular foundation models are also a narrower category than general intelligence, which is the source of both their defensibility and their ceiling. If the frontier labs decide that structured data prediction is worth doing properly, the technical distance is not obviously insurmountable.

But the counterfactual is the point. Eighteen months ago this was a Nature paper. Today it is a funded frontier lab in Baden-Württemberg with open weights and a route into hundreds of thousands of corporate systems. That has not happened in Europe before, and the industry's near-total lack of interest in the news is a reasonable proxy for how underpriced tabular foundation models still are.